Announcement

Collapse
No announcement yet.
X
  • Filter
  • Time
  • Show
Clear All
new posts

  • xtfevd: extremely wide confidence intervals

    Hi,
    I've been using the xtfevd.ado package on a dataset as I have a number of rarely changing variables. I am getting incredibly large confidence intervals. To test the generalizability of this error (at least I suspect I'm doing something wrong) I made a dataset with variables that I was sure to reveal strong relationships. I tested this dataset and model with OLS, RE, FE, xthtaylor and finally also with xtfevd.
    For most variables I get very narrow confidence intervals (P values of around 0.00), but these jump to 0.9 or above when using xtfevd. I notice that there's an incredibly high nr of degress of freedom too, but I'm lost as what I'm doing wrong.
    Output of this testing provided in the attachment, including explanation of variables (at the bottom of this attachment).

    Any suggestions would be greatly appreciated.
    Thanks,

    Jorrit

    Attached Files
    Last edited by Jorrit Gosens; 22 Jan 2015, 09:08. Reason: edited because of error messages with uploads...

  • #2
    Jorrit, apparently xtfevd.ado is a user-written command. The FAQ asks that the source of non-Stata commands be included in the problem description. Attempts to learn about xtfevd using search xtfevd yield no results. Based on the output you supplied "fevd" means "fixed effects with vector decomposition", which I point out to other readers as differing from "forecast error variance decomposition" also abbreviated "fevd" elsewhere in Stata. Google yields a link here to the xtfevd ado and help files, for those interested. The help file does identify the authors, who may be able to assist. On Statlist you are probably limited in the pool of those likely to be able to help you to folk familiar with xtfevd.ado. That's likely a small pool, given that it's not easily discoverable through Stata.

    With that said, a casual glance at your output shows that xtfevd yielded an increase in R-squared to something very near one, which suggests to me the possibility of overfitting in the xtfevd formulation.

    Comment


    • #3
      There is strong advice against using the "Fixed Effects Vector Decomposition" (FEVD).

      First, as you can confirm with the regression output you provided, the coefficients for time-varying regressors are exactly the same as with the standard fixed effects estimator (xtreg, fe). Second, standard error estimates obtained with xtfevd are inconsistent. Third, the xtfevd estimator for the coefficients of time-invariant regressors is inconsistent if those regressors are endogenous.

      For an extensive critique on the FEVD, you might want to read in particular:
      Breusch et al. (2011): "On the Fixed-Effects Vector Decomposition", Political Analysis (19), 123-134;
      Greene (2011): "Fixed Effects Vector Decomposition: A Magical Solution to the Problem of Time-Invariant Variables in Fixed Effects Models?", Political Analysis (19), 135-146.

      Also, the following earlier Statalist discussions on time-invariant regressors in panel data models might be of interest:
      http://www.statalist.org/forums/foru...-effects-model
      http://www.statalist.org/forums/foru...riant-variable
      http://www.statalist.org/forums/foru...-data-analysis
      Last edited by Sebastian Kripfganz; 22 Jan 2015, 10:22.
      https://twitter.com/Kripfganz

      Comment


      • #4
        Thanks for the responses so far.
        the xtfevd package is introduced in the paper Plümper, Thomas, and Vera E. Troeger. "Efficient estimation of time-invariant and rarely changing variables in finite sample panel analyses with unit fixed effects." Political Analysis 15.2 (2007): 124-139 (link) and the ado file is available here.

        I'm aware of the critical article you mention. The original authors have also responded to this criticism: Plümper, Thomas, and Vera E. Troeger. "Fixed-effects vector decomposition: properties, reliability, and instruments." Political Analysis 19.2 (2011): link
        Another critical paper is
        Breusch, Trevor, et al. "On the fixed-effects vector decomposition." Political Analysis 19.2 (2011): 123-134. (Link)

        Which of the authors is more correct I cant really say. The mathematical proof is these papers is quite over my head.

        What I do know is that the xtfevd is suggested as most powerful in dealing with rarely changing variables, and offers the benefit over FE in that entirely time-invariant variables may be included. That matches quite well with a penal of countries that includes a number of physical descriptions as well as a number of policy and economical dummies or count variables that change only very infrequently.

        Note, for example, that the dummy variable
        ContCommRE in the output provided has very different values in xtfevd and FE estimations, even if the other variables are exactly the same.
        Edit: this different value for the dummy occurs when i include it in the invariant() option of xtfevd, same value if its not included there. I've tried xthtaylor as well, and that allows for the perfectly time-invariant (e.g., physical country properties) to be included, but none of the other models mentioned have specific options for rarely changing variables such as dummies. The (potentially false) efficiency in standard errors is one issue, the level of the reported estimator is another. I'm interested in the strength of the influence of individual variables, including the invariant and rarely changing ones.


        Last edited by Jorrit Gosens; 23 Jan 2015, 01:21.

        Comment


        • #5
          I am aware of the original paper by Plümper and Troeger. Classifying rarely changing variables as time-invariant in the FEVD procedure is econometrically anything but clean. As can be seen from your xtreg, fe output, the (rare) time-variation of ContCommRE has predictive power. Leaving it out at the first step of the FEVD procedure then potentially induces an omitted variable bias.

          Another point that strikes me, from your explanation of the variables (in your attached txt-file) I understand that ContCommRE is essentially a function of your dependent variable and therefore is inherently endogenous.
          https://twitter.com/Kripfganz

          Comment


          • #6
            hello,

            Please somebody can help me with the IV-FEVD stata command from Plumper and Troeger (2011)

            Thanks

            Comment


            • #7
              As mentioned above, I cannot recommend to use the IV-FEVD procedure. This does not improve in any respect over the two-stage procedure that you can implement with my xtseqreg command, and you would run the risk of getting incorrect standard errors when using the xtfevd command.
              https://twitter.com/Kripfganz

              Comment


              • #8
                My problem is that. I want to use the static and Dynamic model in my work. For the dynamic model don't have problem because I use your xtseqreg. but for static model, I don't have have just this suggestion command ( IV-FEVD) or FEF-IV estimators from Zhou (2017) this last procedure does not my data as invariant data (legal origin for exemple). I am so trouble
                Attached Files

                Comment


                • #9
                  You can use my xtseqreg command for static models as well. Just remove the lagged dependent variable (and the respective instruments for it).
                  https://twitter.com/Kripfganz

                  Comment

                  Working...
                  X